ZeroGPU × Dappier: real-time classification at scale
A purpose-built ZLM edge model replaced a general-purpose model + RAG pipeline — keeping accuracy while cutting latency and cost across Dappier's live publisher network.
Dappier brings AI answers — and a new ad surface
Dappier distributes a fully-trained "Ask AI" stack to media publishers, who embed it on their sites. In return, Dappier holds exclusive monetization rights to that conversational surface — serving native ads backed by direct relationships with major blue-chip advertisers.
Before ZeroGPU: slow and expensive
A general-purpose nano model (GPT-5.4-nano) paired with a Pinecone RAG lookup handled IAB classification and intent-signal extraction.
Model inference dominated per req ~1,800–2,000 ms; the RAG lookup added ~100 ms more (query + network)
- RAG added infrastructure complexity and per-query cost
- ~2,000 ms latency was too high for real-time ad surfaces
- Token spend scaled linearly with traffic
After ZeroGPU: right-sized ZLM models on the edge
Two purpose-built ZLM models — trained on IAB Content Taxonomy 2.2 + Audience Taxonomy 1.1, fine-tuned and proprietary to ZeroGPU. The RAG lookup was eliminated entirely.
| Metric | Before | After | Δ |
|---|---|---|---|
| Input tokens | 2,000–3,000 | 400 | ~6× fewer |
| Output tokens | 700 | 500 | lower |
| Latency | ~1,800–2,000 ms | 48 ms p50 | ~10× faster |
| Cost / 1M req | $1,375 | $216 | ~6× cheaper |
| Accuracy | baseline | maintained | ✓ |
Model tokens only; RAG infrastructure also eliminated.
Not every task needs a frontier model.
We've been using ZeroGPU's ZLM models at Dappier for IAB classification and signal extraction across our AI interactions, and it's been a meaningful upgrade. We've seen roughly 10x faster latency and significant cost savings compared to using frontier models for the same tasks.